Soil Sample Search in Partially Observable Environments

TL;DR

The paper addresses autonomous soil sampling in unknown outdoor environments with amorphous target regions and limited visibility. It introduces a centroid-based heuristic that guides navigation toward the largest observed soil contour, augmented by a height-control dimension to modulate the camera field of view, and balances exploration versus descent with a sigmoid probability on height. In extensive simulations, the Lissajous-based heuristic consistently outperforms baseline patterns in terms of steps, distance traveled, and robust success across varying soil abundance and visibility. The approach is simple to implement, adaptable to UAVs/UGVs, and yields practical benefits for efficient soil sampling, with potential extensions to frontier exploration, reinforcement learning, and hardware validation.

Abstract

To work in unknown outdoor environments, autonomous sampling machines need the ability to target samples despite limited visibility and robotic arm reach distance. We design a heuristic guided search method to speed up the search process and more efficiently localize the approximate center of soil regions. Through simulation experiments, we assess the effectiveness of the proposed algorithm and discover superior performance in terms of speed, distance traveled, and success rate compared to naive baselines.

Figures (8)

• Figure 1: In this illustration, the contour of the spotted soil samples are shown as the blue region in the rightmost image. The centroid, indicated by the white dot, is determined from the contour and used to define the next waypoint and action, shown by the red arrow.
• Figure 2: In this example, a simulation environment with an amorphous soil distribution is shown. The dark regions represent soil, that can be sampled, and the lighter regions represent all other materials.
• Figure 3: To artificially impair the robot's visibility, we add noise, weighted by $\alpha$, to the ground truth view. The blended image is then passed through a threshold to obtain a binary valued image. In this example, parts of the original soil region are hidden from the agent by noise.
• Figure 4: The $y$ values correspond to the percentage of pixels that are misclassified. In a low visibility environment, nearly all of the soil samples would be missed at a zoom size of above 2x, corresponding to around 40% between $h_{\max{}}$ and $h_{\min{}}$, and around 1-2% of the remaining region would be detected as soil. On the other hand, in a high visibility environment, only a small portion of the soil samples will be missed in most cases, while the false positive rate remains at a similar level.
• Figure 5: Among the four tested method, the Lissajous-based heuristic method outperforms all other methods in terms of step count and distance traveled while maintaining a near 100% success rate across different soil abundance levels.